Keep and Extent: Unified Knowledge Embedding for Few-Shot Image Generation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Training Generative Adversarial Networks (GANs) with limited data is challenging.
• Existing methods struggle with selecting compatible knowledge from source domains for target domains.
• Overfitting to scarce target data is a common issue in few-shot image generation.

Purpose of the Study:

• To propose a unified learning paradigm for improved knowledge transfer in few-shot GAN training.
• To address the limitations of manual selection of compatible knowledge.
• To enhance the performance of GANs in low-data regimes.

Main Methods:

• Introduced a novel Keep and Extend (KAE) learning paradigm.
• Orthogonally decomposed the latent space of GANs.
• Utilized resting latent directions to extend target subspaces while preserving source subspaces.

Main Results:

• The KAE method automatically transfers compatible knowledge without manual selection.
• Experimental results demonstrate the superiority of the proposed method on benchmark datasets.
• Achieved better performance in few-shot image generation compared to existing approaches.

Conclusions:

• The KAE paradigm offers an effective solution for few-shot GAN training.
• Orthogonal decomposition of the latent space facilitates robust knowledge transfer.
• This approach mitigates overfitting and improves generative capabilities with limited data.